01Is the GPU actually mine?

“Dedicated” is used to mean at least three different things: a whole physical GPU reserved to you, a time-sliced share of one, or a MIG partition. The first gives predictable throughput. The others give you a number on an invoice and variance in your training times that you will spend a week failing to explain.

Ask for the specific arrangement, then verify it yourself rather than taking the answer on trust. Our post on what dedicated GPU actually means, and how to check covers the commands.

Our position: no shared tenancy on your allocated GPU, on any plan.

02Will my job even fit?

Providers sell compute; jobs die on memory. Before comparing anyone on price, work out the VRAM your actual model, batch size and sequence length require — because a cheaper card that cannot hold your training step is not cheaper, it is unusable.

The arithmetic is in sizing VRAM for training and serving. Bring the resulting number to every provider conversation. It converts a vague comparison into a yes or no.

03What is the billing unit?

Per-second, per-hour and per-month are not variations on a theme — they reward completely different usage patterns. Hourly billing suits spiky work and punishes idle reservations. Monthly billing suits steady work and punishes idleness absolutely.

Work out your own utilisation honestly, then find the break-even. We published ours, including the rows where we lose: what a fine-tune costs when the GPU is billed by the month. A provider unwilling to show you the arithmetic that makes them look bad is telling you something.

04What is metered, and what is quoted?

The headline price is rarely the invoice. The usual surprises are storage beyond an included allowance, egress bandwidth, snapshot retention, and a second instance you spun up for a week and forgot.

Ask which components are metered, what is included, and whether overages are quoted in advance or simply appear. Ours: 500 GB NVMe on Starter and 2 TB on Professional, egress on fair use, and anything beyond quoted before you incur it rather than billed silently. The full table is on the pricing page.

05Where is the hardware, and whose law applies?

Two distinct questions. Physical location determines latency and which government can compel disclosure. The contracting entity determines which courts hear a dispute and what recourse you actually have.

Ask for a named country and facility type, and for the legal entity on the contract. “Global infrastructure” is not an answer. Then ask separately whether account and billing data stays in the same place as the workload — it usually does not, ours included, and a provider who claims otherwise has probably not checked their own payment processor. We go through this in where your prompts actually go.

06What happens when something breaks?

Published support hours matter more than a published uptime figure, because the hours tell you when a human will actually read your message. Ask for the hours, the target for a first human reply, and the escalation path for a production outage.

Ours, plainly: Monday to Friday, 10:00 to 19:00 IST, with a target of a reply within one business day, and a phone number on the contact page. Enterprise arrangements include 24/7 support and an account manager. If you need follow-the-sun coverage on a standard plan, we do not offer it and you should weight that accordingly.

07How do I leave?

Ask this before you sign, not when you want out. Four specifics: is there a setup fee, is there a minimum term, how much notice does cancellation need, and how long after termination is your data deleted?

Ours: no setup fee, no lock-in, and the first month carries a 14-day full refund so you can test with your real workload rather than an estimate. Retention periods after termination are set out in the Privacy Policy — access and security logs for 90 days, usage and billing telemetry for 13 months, support correspondence for 24 months.

08Where we are the wrong answer

Four cases where we would tell you to look elsewhere, so you do not discover them a month in:

• Genuinely spiky workloads. One short fine-tune a quarter, or a burst around a launch and nothing between. Per-second billing elsewhere will beat a monthly reservation comfortably.
• 24/7 support on a standard plan. Our staffed hours are Monday to Friday, business hours IST. If an outage at 03:00 needs a human immediately, that is an Enterprise arrangement or another provider.
• Multi-region redundancy. Our hardware is in Mumbai. If your architecture requires failover across regions, one facility does not provide it.
• A specific certification. Our Privacy Policy states that we claim no certification or audit report we have not actually obtained. If procurement requires a named one, ask first.

If your workload does fit — steady utilisation, memory that suits a 16 GB or 40 GB card, and a preference for inference that does not leave your tenancy — describe it and we will size it: contact@vijaycloud.com.